Digital twins in industry: what they really deliver, and what they cost

The digital twin is everywhere: in conferences, vendor brochures and executive roadmaps. On paper, you replicate your plant inside the computer, you see everything and you anticipate everything. In reality, many sites have paid dearly for a gorgeous 3D model that nobody opens a year later. The gap between the two does not come from technology. It comes from a question almost everyone skips: what decision is this twin meant to support?

Three very different things under one word

The term covers three levels that get confused, and that confusion explains a good part of the disappointments.

The digital model is a static representation: a drawing, a simulation, a mock-up. It does not move when the installation moves. Useful to design, not enough to operate.

The digital shadow adds a flow: the real data from the equipment feeds back into the model, which reflects the current state. You see what is happening, with a slight delay. That alone is valuable, and it is where many honest projects stop.

The digital twin closes the loop: the model receives the data, interprets it, and sends something back toward the decision or the installation, a setting, an alert, a recommendation. The difference between the three is not the beauty of the rendering. It is the presence, or absence, of a data loop that lives.

Why so many twins end up as screensavers

Three traps recur, and none of them is technical.

The cathedral. People want to model the whole plant before knowing what it will be used for. The project grows, costs, takes months, and produces an impressive object that nothing connects to a daily decision. A twin that answers everything answers, in practice, nothing.

Data that does not flow. A twin lives on its data. If sensors are missing, if readings are manual and irregular, if systems do not talk to each other, the model freezes and drifts from reality. Six months later, it describes a plant that no longer exists.

The orphan twin. Nobody is tasked with keeping it current. Yet a twin degrades like a piece of equipment: without upkeep, it lies. The question of the owner, the person accountable for it, comes before the question of the software.

Where AI genuinely changes things

A twin without intelligence is a slightly rich dashboard: it shows. It is AI that turns it from a mirror into a decision tool, and that is exactly the subject at hand, AI applied to industry, not 3D for its own sake.

Three uses deliver on their promise. Anomaly detection: the AI learns the normal behaviour of the process as the twin describes it, and flags the deviation before it becomes a breakdown. Projection: from the measured trends, it estimates where the equipment is heading, a subject we cover in detail for estimating the remaining useful life of an asset. Scenario simulation: because the twin holds a model of your real installation, you can ask it "what happens if", raise a throughput or change a parameter, and get an answer grounded in your data rather than in an intuition.

One does not work without the other. The twin without AI lacks exploitation; AI without the twin lacks context. Together, they let you test a decision before making it on the real installation.

The cost question, without dodging it

A digital twin is paid for in several places, and the software is not the heaviest. There is the platform licence, of course. But above all there is integration, getting data that often does not talk to flow and connect; data quality, without which the rest collapses; and the upkeep of the twin itself, that recurring cost nobody prices at the start and which alone decides whether the project survives.

That is why a twin is justified by a decision that is worth a lot: an investment trade-off, a production bottleneck, a critical asset whose downtime costs a fortune. For a low-stakes decision, a dashboard is enough, and it costs a hundred times less. Deciding on a twin is first an honest business case, not a choice of tool.

Build around a decision, and keep control

The thread of a responsible twin is simple: it prepares the decision, it does not make it for you. A model, however faithful, remains an approximation of reality. Its confidence must never exceed the quality of its data, and the human keeps control of what the twin recommends, exactly as for any industrial use of AI where the decision carries weight on safety or investment. A twin that operated a critical asset on its own would cross a line the stakes do not allow.

That requirement is not a brake, it is what makes the project solid. A twin built around a clear decision, fed by data you control, held by someone accountable for it, delivers real value. A twin built to impress a board ends up as a demonstration you wheel out once a year.

Where to start, without a big project

You do not need to model everything to begin. Take a process you know well, tie it to a decision that recurs, and use the data you already have. If the digital shadow of that single process helps you decide better, you will know the full twin is worth it. If not, you will have saved yourself a cathedral. The logic is that of any first AI project framed over ninety days: a narrow scope, a proof, then you extend.

The essentials

A digital twin is neither a fad to flee nor a magic wand. It is a decision tool worth exactly what its data and the question it answers are worth. Model a process, not a plant. Tie it to a decision that matters. Let AI exploit it and the human decide. On those conditions, it delivers a great deal. Without them, it costs a lot and serves little.